Blood Bots

The discourse on the marriage of artificial intelligence and pathology is hardly fresh at this point. If you’re a keen reader of The Pathologist, you’ll be familiar with the many articles we’ve published on the matter – including a spicy (if I do say so myself) discussion on the ethics of AI in general. Self-publicity aside, there are certain AI areas that are lacking their fair share of coverage. Demonstrating real-life applications of AI in clinical settings, for instance, is still a fledgling branch of research. And though they are starting to increase in number, it feels like each new analysis of AI’s importance in the diagnostic pipeline takes us one step closer towards a new era of healthcare.

One such study is offered by Carsten Marr, Director of the Institute of AI for Health at Helmholtz Munich, Germany, and his team, who used the increasing availability and capabilities of AI to support diagnostic decision making in hematopathology (1). In this case, the aim was to assist pathologists in spotting acute myeloid leukemia (AML). I spoke with Marr to find more about the study and the implications it could have on hemepath as a whole. 

First, could you broadly explain your work and how you arrived there?
 

Today, my team develops machine learning and mechanistic models to boost diagnostics and understand stem cell kinetics. How did I get here? I studied physics in Munich, did my PhD in network biology, and transitioned from bioinformatics to clinical data analysis and health AI during my time as postdoc and group leader. Later, I got involved in computational hematopathology as it was quite clear to me – based on my discussions with clinicians and pathologists – that the amount of data in histopathological slides was a perfect use case for health AI.

What work have you been doing in blood diagnostics?
 

We trained several models with different blood datasets to address questions such as: Can we automatically discriminate blood cells? Can we predict the disease or subdisease? The genetic alteration? From blood, bone marrow smears, and histological sections?

If so, with what accuracy? And based on which morphological features?

How were you able to train an AI system to detect different AML subtypes?
 

With the help of a large enough data set, sufficient computing power, and dedicated students – ha!

AI worked well in cases of leukemia. Will this approach work for other blood diseases? Do you expect difficulties in other areas?
 

I expect it to work wherever pathologists and cytologists can do the job. What these experts train into their brain over years of education, we train into the parameters of our artificial neural networks. It becomes particularly interesting when we can identify diseases where experts fail.

Your system’s classification accuracy was comparable to human experts. Is this a huge deal for diagnostics or should we hold back our excitement?
 

I think it’s great. To make a difference in the clinics or the lab, we have to make sure algorithms work with real-world data (and not particularly selected datasets), can deal with data from different labs, and are efficiently integrated in the existing workflows.

Your AI was able to detect even very rare forms of cancer, where less data is available. Are there any pitfalls?
 

The more difficult the problem, the more data we need. However, if some cancer cells are very different from all others’ data, models can also work with a limited number of examples (2).

How do we bring these AI models into clinical settings?
 

Optimize workflows. Make sure it makes the experts’ lives easier. Evaluate workflows with  algorithms against traditional ones and measure time and performance. Plus, we need hardware companies to join in.

You’ve spoken about a healthcare data “explosion” in the near future. Is this purely a positive thing or are there any negatives to this glut of info?
 

I think there is really a great deal we can learn, once we are able to analyze our health data. Of course, we need to make sure that individuals have control over their data, which they currently do not have. I suggest moving forward with caution but note that we should definitely move now.